The Hong Kong Monetary Authority announced its ART Framework on October 15 to support AI adoption in banking, with accountability, resilience and trust at its core.
The announcement addresses a change brought by increasingly autonomous systems: as AI gains access to more data, tools and workflows, governance needs to cover its actions and the ability to intervene when problems arise.
Guidance issued, agent testing continues
The HKMA’s measures include issuing updated supervisory guidance for responsible AI adoption and working with Cyberport and the University of Hong Kong on an AI training data sharing platform, with data quality assurance as a core design principle.
For agentic AI, the HKMA will continue application testing through GenA.I. Sandbox++ and plans a dedicated governance discussion paper. Areas to explore include intent-anchored accountability, a registry of material AI agents and effective intervention mechanisms.
These measures are at different stages. The supervisory guidance has been issued, while dedicated agentic governance proposals remain subject to discussion. The announcement’s directions should not all be treated as a single set of rules already implemented.
From generating answers to executing tasks
The HKMA says more than 60 percent of use cases in the second cohort of its generative AI sandbox explored agentic AI, including automated planning and execution of multiple steps within defined boundaries and controls.
That proportion describes this cohort’s use cases. It cannot be used to infer deployment across the entire banking industry.
Generating an explanation and then acting on it in a software system raise different management questions. If an assistant can initiate further processing after organizing information, for example, its permitted data access, actions and human decision points need to be clear. This illustrates a governance issue rather than a specific bank case disclosed in the announcement.
Intervention belongs in workflow design
AIYXL Perspective: As AI participates in execution, accountability needs to reach individual actions: who set the goal, who granted permissions, what informed the system’s choices, and how people can detect and stop behavior that departs from the goal.
Application teams should also test abnormal conditions. Whether a system pauses and requests help when information is missing, a tool returns an error or a task’s goal changes can reveal more about workflow reliability than a successful routine demonstration.
The ART Framework’s direction brings experimentation, data quality and human oversight into the development of banking AI applications.